Bounds on Mahalanobis norms and their applications (Q1369288)

From MaRDI portal
Revision as of 19:23, 27 May 2024 by ReferenceBot (talk | contribs) (‎Changed an Item)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
scientific article
Language Label Description Also known as
English
Bounds on Mahalanobis norms and their applications
scientific article

    Statements

    Bounds on Mahalanobis norms and their applications (English)
    0 references
    20 October 1997
    0 references
    Let \({\mathcal P}_n \in\mathbb{R}^{n \times n}\) be the set of positive definite matrices and \(|x|_M= \sqrt{x^\top Mx}\), \(M\in{\mathcal P}_n\). For \(A,B\in {\mathcal P}_n\) denote \(D_\gamma =\text{diag} (\gamma_1, \dots, \gamma_n)\), where \((\gamma_i,q_i) \in\mathbb{R}_+ \times \mathbb{R}^n\) are the eigenpairs of the generalized eigenvalue problem \(Aq= \gamma Bq\). Let \(B^{-1/2} AB^{-1/2} =PD_\gamma P^\top\), where \(P\) is an orthogonal matrix. Set \(\alpha_i= \min \{1,\gamma_i\}\), \(\beta_i= \max \{1,\gamma_i\}\) and \(A\vee B=B^{1/2} PD_\alpha P^\top B^{1/2}\), \(A\wedge B=B^{1/2} PD_\beta P^\top B^{1/2}\). Suppose that \(\gamma_1\geq \cdots \geq\gamma_r=1 =\cdots= \gamma_{r+s} >\gamma_{r+s+1} \geq \cdots \geq \gamma_n >0\) (at least one of the integers \(r,s\) exists) and set \(R_1=\text{span} (q_1, \dots, q_r)\), \(R_2= \text{span} (q_{r+1}, \dots, q_{r+s})\), \(R_3= \text{span} (q_{r+s+1}, \dots, q_n)\). Denote \(\lambda(x) =|x|_A/ |x|_B\), \(0\neq x\in \mathbb{R}^n\) and \(\eta_m(x) =|x|_{A\vee B}\), \(\eta_M(x)=|x|_{A\wedge B}\). It is shown that: (i) \(\sqrt {\gamma_n} \leq\lambda (x)\leq \sqrt {\gamma_1}\) for \(x\in\mathbb{R}^n\); (ii) \(\sqrt {\gamma_r} \leq \lambda (x)\) for \(x\in R_1\); (iii) \(\lambda (x)=1\) for \(x\in R_2\) and (iv) \(\lambda (x)\leq \sqrt {\gamma_{r+s+1}}\) for \(x\in R_3\). The following bounds are also derived: (j) \(\eta_m (x) \leq |x|_A\), \(|x|_B \leq\eta_M(x)\) for \(x\in \mathbb{R}^n\); (jj) \(\eta_m(x) =|x |_B\) and \(|x|_A =\eta_M(x)\) for \(x\in R_1\); (jjj) \(\eta_m(x) =|x|_B =|x|_A =\eta_M (x)\) for \(x\in R_2\) and (jv) \(\eta_m(x) =|x|_A\) and \(|x|_B = \eta_M(x)\) for \(x\in R_3\).
    0 references
    0 references
    bounds
    0 references
    Mahalanobis norms
    0 references
    positive definite matrices
    0 references
    generalized eigenvalue problem
    0 references
    0 references